Loops HouseChance

NEAR HACKATHON
Links
Team
1 member- LOOwner
Lorex
Overview
ChanceNear is the self-improvement stack for AI agents on NEAR Intents. Agents with wallets fail in ways a spending limit can't see: they follow instructions hidden in an invoice, convert 25 USDC with the wrong decimals, sell when the user said buy, or pay a lookalike account. A typical agent guard only blocks; ChanceNear blocks and then makes the agent better. Layer 1, Catch: before a swap, transfer, withdrawal or limit order runs, it checks the exact Agent API body against the user's own words (amounts, decimals, shares of a balance, tokens, chain, side, recipient), traces whether the recipient or amount came from injected text, reads NEAR's own numbers (GET /v1/tokens, Agent API dry runs, live 1Click quotes) and asks a NEAR AI Cloud model in a TEE, verifying the enclave's signature over the exact request and response. A model can add doubt but never lift a deterministic block. Layers 2-4, Learn, Remember, Measure: every BLOCK is stored in Postgres as a lesson in that agent's playbook, handed back in every verdict, as playbook.md (a Claude Code SessionStart hook) and over MCP; lessons are credited when they prevent a repeat, and rehearsals graph mistakes falling round by round. Layer 5, Train (optional): SFT/DPO export for open-weight agents such as Kimi K3, Qwen or DeepSeek. Allowed actions are built to run through the Agent API with reference=sha256:, so NEAR's attested receipts name the verdict that approved them; that live step switches on once the Agent API accepts our partner key, and until then swaps are grounded in live 1Click quotes. Tracks (confirmed): security, privacy